xref: /openbmc/linux/net/sunrpc/xprtrdma/verbs.c (revision 48c926cd)
1 /*
2  * Copyright (c) 2003-2007 Network Appliance, Inc. All rights reserved.
3  *
4  * This software is available to you under a choice of one of two
5  * licenses.  You may choose to be licensed under the terms of the GNU
6  * General Public License (GPL) Version 2, available from the file
7  * COPYING in the main directory of this source tree, or the BSD-type
8  * license below:
9  *
10  * Redistribution and use in source and binary forms, with or without
11  * modification, are permitted provided that the following conditions
12  * are met:
13  *
14  *      Redistributions of source code must retain the above copyright
15  *      notice, this list of conditions and the following disclaimer.
16  *
17  *      Redistributions in binary form must reproduce the above
18  *      copyright notice, this list of conditions and the following
19  *      disclaimer in the documentation and/or other materials provided
20  *      with the distribution.
21  *
22  *      Neither the name of the Network Appliance, Inc. nor the names of
23  *      its contributors may be used to endorse or promote products
24  *      derived from this software without specific prior written
25  *      permission.
26  *
27  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
28  * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
29  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
30  * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
31  * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
32  * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
33  * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
34  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
35  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
36  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
37  * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
38  */
39 
40 /*
41  * verbs.c
42  *
43  * Encapsulates the major functions managing:
44  *  o adapters
45  *  o endpoints
46  *  o connections
47  *  o buffer memory
48  */
49 
50 #include <linux/interrupt.h>
51 #include <linux/slab.h>
52 #include <linux/prefetch.h>
53 #include <linux/sunrpc/addr.h>
54 #include <linux/sunrpc/svc_rdma.h>
55 #include <asm/bitops.h>
56 
57 #include <rdma/ib_cm.h>
58 
59 #include "xprt_rdma.h"
60 
61 /*
62  * Globals/Macros
63  */
64 
65 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
66 # define RPCDBG_FACILITY	RPCDBG_TRANS
67 #endif
68 
69 /*
70  * internal functions
71  */
72 static void rpcrdma_create_mrs(struct rpcrdma_xprt *r_xprt);
73 static void rpcrdma_destroy_mrs(struct rpcrdma_buffer *buf);
74 static void rpcrdma_dma_unmap_regbuf(struct rpcrdma_regbuf *rb);
75 
76 static struct workqueue_struct *rpcrdma_receive_wq __read_mostly;
77 
78 int
79 rpcrdma_alloc_wq(void)
80 {
81 	struct workqueue_struct *recv_wq;
82 
83 	recv_wq = alloc_workqueue("xprtrdma_receive",
84 				  WQ_MEM_RECLAIM | WQ_UNBOUND | WQ_HIGHPRI,
85 				  0);
86 	if (!recv_wq)
87 		return -ENOMEM;
88 
89 	rpcrdma_receive_wq = recv_wq;
90 	return 0;
91 }
92 
93 void
94 rpcrdma_destroy_wq(void)
95 {
96 	struct workqueue_struct *wq;
97 
98 	if (rpcrdma_receive_wq) {
99 		wq = rpcrdma_receive_wq;
100 		rpcrdma_receive_wq = NULL;
101 		destroy_workqueue(wq);
102 	}
103 }
104 
105 static void
106 rpcrdma_qp_async_error_upcall(struct ib_event *event, void *context)
107 {
108 	struct rpcrdma_ep *ep = context;
109 
110 	pr_err("rpcrdma: %s on device %s ep %p\n",
111 	       ib_event_msg(event->event), event->device->name, context);
112 
113 	if (ep->rep_connected == 1) {
114 		ep->rep_connected = -EIO;
115 		rpcrdma_conn_func(ep);
116 		wake_up_all(&ep->rep_connect_wait);
117 	}
118 }
119 
120 /**
121  * rpcrdma_wc_send - Invoked by RDMA provider for each polled Send WC
122  * @cq:	completion queue (ignored)
123  * @wc:	completed WR
124  *
125  */
126 static void
127 rpcrdma_wc_send(struct ib_cq *cq, struct ib_wc *wc)
128 {
129 	/* WARNING: Only wr_cqe and status are reliable at this point */
130 	if (wc->status != IB_WC_SUCCESS && wc->status != IB_WC_WR_FLUSH_ERR)
131 		pr_err("rpcrdma: Send: %s (%u/0x%x)\n",
132 		       ib_wc_status_msg(wc->status),
133 		       wc->status, wc->vendor_err);
134 }
135 
136 /* Perform basic sanity checking to avoid using garbage
137  * to update the credit grant value.
138  */
139 static void
140 rpcrdma_update_granted_credits(struct rpcrdma_rep *rep)
141 {
142 	struct rpcrdma_buffer *buffer = &rep->rr_rxprt->rx_buf;
143 	__be32 *p = rep->rr_rdmabuf->rg_base;
144 	u32 credits;
145 
146 	credits = be32_to_cpup(p + 2);
147 	if (credits == 0)
148 		credits = 1;	/* don't deadlock */
149 	else if (credits > buffer->rb_max_requests)
150 		credits = buffer->rb_max_requests;
151 
152 	atomic_set(&buffer->rb_credits, credits);
153 }
154 
155 /**
156  * rpcrdma_wc_receive - Invoked by RDMA provider for each polled Receive WC
157  * @cq:	completion queue (ignored)
158  * @wc:	completed WR
159  *
160  */
161 static void
162 rpcrdma_wc_receive(struct ib_cq *cq, struct ib_wc *wc)
163 {
164 	struct ib_cqe *cqe = wc->wr_cqe;
165 	struct rpcrdma_rep *rep = container_of(cqe, struct rpcrdma_rep,
166 					       rr_cqe);
167 
168 	/* WARNING: Only wr_id and status are reliable at this point */
169 	if (wc->status != IB_WC_SUCCESS)
170 		goto out_fail;
171 
172 	/* status == SUCCESS means all fields in wc are trustworthy */
173 	dprintk("RPC:       %s: rep %p opcode 'recv', length %u: success\n",
174 		__func__, rep, wc->byte_len);
175 
176 	rpcrdma_set_xdrlen(&rep->rr_hdrbuf, wc->byte_len);
177 	rep->rr_wc_flags = wc->wc_flags;
178 	rep->rr_inv_rkey = wc->ex.invalidate_rkey;
179 
180 	ib_dma_sync_single_for_cpu(rdmab_device(rep->rr_rdmabuf),
181 				   rdmab_addr(rep->rr_rdmabuf),
182 				   wc->byte_len, DMA_FROM_DEVICE);
183 
184 	if (wc->byte_len >= RPCRDMA_HDRLEN_ERR)
185 		rpcrdma_update_granted_credits(rep);
186 
187 out_schedule:
188 	queue_work(rpcrdma_receive_wq, &rep->rr_work);
189 	return;
190 
191 out_fail:
192 	if (wc->status != IB_WC_WR_FLUSH_ERR)
193 		pr_err("rpcrdma: Recv: %s (%u/0x%x)\n",
194 		       ib_wc_status_msg(wc->status),
195 		       wc->status, wc->vendor_err);
196 	rpcrdma_set_xdrlen(&rep->rr_hdrbuf, 0);
197 	goto out_schedule;
198 }
199 
200 static void
201 rpcrdma_update_connect_private(struct rpcrdma_xprt *r_xprt,
202 			       struct rdma_conn_param *param)
203 {
204 	struct rpcrdma_create_data_internal *cdata = &r_xprt->rx_data;
205 	const struct rpcrdma_connect_private *pmsg = param->private_data;
206 	unsigned int rsize, wsize;
207 
208 	/* Default settings for RPC-over-RDMA Version One */
209 	r_xprt->rx_ia.ri_reminv_expected = false;
210 	r_xprt->rx_ia.ri_implicit_roundup = xprt_rdma_pad_optimize;
211 	rsize = RPCRDMA_V1_DEF_INLINE_SIZE;
212 	wsize = RPCRDMA_V1_DEF_INLINE_SIZE;
213 
214 	if (pmsg &&
215 	    pmsg->cp_magic == rpcrdma_cmp_magic &&
216 	    pmsg->cp_version == RPCRDMA_CMP_VERSION) {
217 		r_xprt->rx_ia.ri_reminv_expected = true;
218 		r_xprt->rx_ia.ri_implicit_roundup = true;
219 		rsize = rpcrdma_decode_buffer_size(pmsg->cp_send_size);
220 		wsize = rpcrdma_decode_buffer_size(pmsg->cp_recv_size);
221 	}
222 
223 	if (rsize < cdata->inline_rsize)
224 		cdata->inline_rsize = rsize;
225 	if (wsize < cdata->inline_wsize)
226 		cdata->inline_wsize = wsize;
227 	dprintk("RPC:       %s: max send %u, max recv %u\n",
228 		__func__, cdata->inline_wsize, cdata->inline_rsize);
229 	rpcrdma_set_max_header_sizes(r_xprt);
230 }
231 
232 static int
233 rpcrdma_conn_upcall(struct rdma_cm_id *id, struct rdma_cm_event *event)
234 {
235 	struct rpcrdma_xprt *xprt = id->context;
236 	struct rpcrdma_ia *ia = &xprt->rx_ia;
237 	struct rpcrdma_ep *ep = &xprt->rx_ep;
238 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
239 	struct sockaddr *sap = (struct sockaddr *)&ep->rep_remote_addr;
240 #endif
241 	int connstate = 0;
242 
243 	switch (event->event) {
244 	case RDMA_CM_EVENT_ADDR_RESOLVED:
245 	case RDMA_CM_EVENT_ROUTE_RESOLVED:
246 		ia->ri_async_rc = 0;
247 		complete(&ia->ri_done);
248 		break;
249 	case RDMA_CM_EVENT_ADDR_ERROR:
250 		ia->ri_async_rc = -EHOSTUNREACH;
251 		dprintk("RPC:       %s: CM address resolution error, ep 0x%p\n",
252 			__func__, ep);
253 		complete(&ia->ri_done);
254 		break;
255 	case RDMA_CM_EVENT_ROUTE_ERROR:
256 		ia->ri_async_rc = -ENETUNREACH;
257 		dprintk("RPC:       %s: CM route resolution error, ep 0x%p\n",
258 			__func__, ep);
259 		complete(&ia->ri_done);
260 		break;
261 	case RDMA_CM_EVENT_DEVICE_REMOVAL:
262 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
263 		pr_info("rpcrdma: removing device %s for %pIS:%u\n",
264 			ia->ri_device->name,
265 			sap, rpc_get_port(sap));
266 #endif
267 		set_bit(RPCRDMA_IAF_REMOVING, &ia->ri_flags);
268 		ep->rep_connected = -ENODEV;
269 		xprt_force_disconnect(&xprt->rx_xprt);
270 		wait_for_completion(&ia->ri_remove_done);
271 
272 		ia->ri_id = NULL;
273 		ia->ri_pd = NULL;
274 		ia->ri_device = NULL;
275 		/* Return 1 to ensure the core destroys the id. */
276 		return 1;
277 	case RDMA_CM_EVENT_ESTABLISHED:
278 		connstate = 1;
279 		rpcrdma_update_connect_private(xprt, &event->param.conn);
280 		goto connected;
281 	case RDMA_CM_EVENT_CONNECT_ERROR:
282 		connstate = -ENOTCONN;
283 		goto connected;
284 	case RDMA_CM_EVENT_UNREACHABLE:
285 		connstate = -ENETDOWN;
286 		goto connected;
287 	case RDMA_CM_EVENT_REJECTED:
288 		dprintk("rpcrdma: connection to %pIS:%u rejected: %s\n",
289 			sap, rpc_get_port(sap),
290 			rdma_reject_msg(id, event->status));
291 		connstate = -ECONNREFUSED;
292 		if (event->status == IB_CM_REJ_STALE_CONN)
293 			connstate = -EAGAIN;
294 		goto connected;
295 	case RDMA_CM_EVENT_DISCONNECTED:
296 		connstate = -ECONNABORTED;
297 connected:
298 		atomic_set(&xprt->rx_buf.rb_credits, 1);
299 		ep->rep_connected = connstate;
300 		rpcrdma_conn_func(ep);
301 		wake_up_all(&ep->rep_connect_wait);
302 		/*FALLTHROUGH*/
303 	default:
304 		dprintk("RPC:       %s: %pIS:%u on %s/%s (ep 0x%p): %s\n",
305 			__func__, sap, rpc_get_port(sap),
306 			ia->ri_device->name, ia->ri_ops->ro_displayname,
307 			ep, rdma_event_msg(event->event));
308 		break;
309 	}
310 
311 	return 0;
312 }
313 
314 static struct rdma_cm_id *
315 rpcrdma_create_id(struct rpcrdma_xprt *xprt,
316 			struct rpcrdma_ia *ia, struct sockaddr *addr)
317 {
318 	unsigned long wtimeout = msecs_to_jiffies(RDMA_RESOLVE_TIMEOUT) + 1;
319 	struct rdma_cm_id *id;
320 	int rc;
321 
322 	init_completion(&ia->ri_done);
323 	init_completion(&ia->ri_remove_done);
324 
325 	id = rdma_create_id(&init_net, rpcrdma_conn_upcall, xprt, RDMA_PS_TCP,
326 			    IB_QPT_RC);
327 	if (IS_ERR(id)) {
328 		rc = PTR_ERR(id);
329 		dprintk("RPC:       %s: rdma_create_id() failed %i\n",
330 			__func__, rc);
331 		return id;
332 	}
333 
334 	ia->ri_async_rc = -ETIMEDOUT;
335 	rc = rdma_resolve_addr(id, NULL, addr, RDMA_RESOLVE_TIMEOUT);
336 	if (rc) {
337 		dprintk("RPC:       %s: rdma_resolve_addr() failed %i\n",
338 			__func__, rc);
339 		goto out;
340 	}
341 	rc = wait_for_completion_interruptible_timeout(&ia->ri_done, wtimeout);
342 	if (rc < 0) {
343 		dprintk("RPC:       %s: wait() exited: %i\n",
344 			__func__, rc);
345 		goto out;
346 	}
347 
348 	rc = ia->ri_async_rc;
349 	if (rc)
350 		goto out;
351 
352 	ia->ri_async_rc = -ETIMEDOUT;
353 	rc = rdma_resolve_route(id, RDMA_RESOLVE_TIMEOUT);
354 	if (rc) {
355 		dprintk("RPC:       %s: rdma_resolve_route() failed %i\n",
356 			__func__, rc);
357 		goto out;
358 	}
359 	rc = wait_for_completion_interruptible_timeout(&ia->ri_done, wtimeout);
360 	if (rc < 0) {
361 		dprintk("RPC:       %s: wait() exited: %i\n",
362 			__func__, rc);
363 		goto out;
364 	}
365 	rc = ia->ri_async_rc;
366 	if (rc)
367 		goto out;
368 
369 	return id;
370 
371 out:
372 	rdma_destroy_id(id);
373 	return ERR_PTR(rc);
374 }
375 
376 /*
377  * Exported functions.
378  */
379 
380 /**
381  * rpcrdma_ia_open - Open and initialize an Interface Adapter.
382  * @xprt: controlling transport
383  * @addr: IP address of remote peer
384  *
385  * Returns 0 on success, negative errno if an appropriate
386  * Interface Adapter could not be found and opened.
387  */
388 int
389 rpcrdma_ia_open(struct rpcrdma_xprt *xprt, struct sockaddr *addr)
390 {
391 	struct rpcrdma_ia *ia = &xprt->rx_ia;
392 	int rc;
393 
394 	ia->ri_id = rpcrdma_create_id(xprt, ia, addr);
395 	if (IS_ERR(ia->ri_id)) {
396 		rc = PTR_ERR(ia->ri_id);
397 		goto out_err;
398 	}
399 	ia->ri_device = ia->ri_id->device;
400 
401 	ia->ri_pd = ib_alloc_pd(ia->ri_device, 0);
402 	if (IS_ERR(ia->ri_pd)) {
403 		rc = PTR_ERR(ia->ri_pd);
404 		pr_err("rpcrdma: ib_alloc_pd() returned %d\n", rc);
405 		goto out_err;
406 	}
407 
408 	switch (xprt_rdma_memreg_strategy) {
409 	case RPCRDMA_FRMR:
410 		if (frwr_is_supported(ia)) {
411 			ia->ri_ops = &rpcrdma_frwr_memreg_ops;
412 			break;
413 		}
414 		/*FALLTHROUGH*/
415 	case RPCRDMA_MTHCAFMR:
416 		if (fmr_is_supported(ia)) {
417 			ia->ri_ops = &rpcrdma_fmr_memreg_ops;
418 			break;
419 		}
420 		/*FALLTHROUGH*/
421 	default:
422 		pr_err("rpcrdma: Device %s does not support memreg mode %d\n",
423 		       ia->ri_device->name, xprt_rdma_memreg_strategy);
424 		rc = -EINVAL;
425 		goto out_err;
426 	}
427 
428 	return 0;
429 
430 out_err:
431 	rpcrdma_ia_close(ia);
432 	return rc;
433 }
434 
435 /**
436  * rpcrdma_ia_remove - Handle device driver unload
437  * @ia: interface adapter being removed
438  *
439  * Divest transport H/W resources associated with this adapter,
440  * but allow it to be restored later.
441  */
442 void
443 rpcrdma_ia_remove(struct rpcrdma_ia *ia)
444 {
445 	struct rpcrdma_xprt *r_xprt = container_of(ia, struct rpcrdma_xprt,
446 						   rx_ia);
447 	struct rpcrdma_ep *ep = &r_xprt->rx_ep;
448 	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
449 	struct rpcrdma_req *req;
450 	struct rpcrdma_rep *rep;
451 
452 	cancel_delayed_work_sync(&buf->rb_refresh_worker);
453 
454 	/* This is similar to rpcrdma_ep_destroy, but:
455 	 * - Don't cancel the connect worker.
456 	 * - Don't call rpcrdma_ep_disconnect, which waits
457 	 *   for another conn upcall, which will deadlock.
458 	 * - rdma_disconnect is unneeded, the underlying
459 	 *   connection is already gone.
460 	 */
461 	if (ia->ri_id->qp) {
462 		ib_drain_qp(ia->ri_id->qp);
463 		rdma_destroy_qp(ia->ri_id);
464 		ia->ri_id->qp = NULL;
465 	}
466 	ib_free_cq(ep->rep_attr.recv_cq);
467 	ib_free_cq(ep->rep_attr.send_cq);
468 
469 	/* The ULP is responsible for ensuring all DMA
470 	 * mappings and MRs are gone.
471 	 */
472 	list_for_each_entry(rep, &buf->rb_recv_bufs, rr_list)
473 		rpcrdma_dma_unmap_regbuf(rep->rr_rdmabuf);
474 	list_for_each_entry(req, &buf->rb_allreqs, rl_all) {
475 		rpcrdma_dma_unmap_regbuf(req->rl_rdmabuf);
476 		rpcrdma_dma_unmap_regbuf(req->rl_sendbuf);
477 		rpcrdma_dma_unmap_regbuf(req->rl_recvbuf);
478 	}
479 	rpcrdma_destroy_mrs(buf);
480 
481 	/* Allow waiters to continue */
482 	complete(&ia->ri_remove_done);
483 }
484 
485 /**
486  * rpcrdma_ia_close - Clean up/close an IA.
487  * @ia: interface adapter to close
488  *
489  */
490 void
491 rpcrdma_ia_close(struct rpcrdma_ia *ia)
492 {
493 	dprintk("RPC:       %s: entering\n", __func__);
494 	if (ia->ri_id != NULL && !IS_ERR(ia->ri_id)) {
495 		if (ia->ri_id->qp)
496 			rdma_destroy_qp(ia->ri_id);
497 		rdma_destroy_id(ia->ri_id);
498 	}
499 	ia->ri_id = NULL;
500 	ia->ri_device = NULL;
501 
502 	/* If the pd is still busy, xprtrdma missed freeing a resource */
503 	if (ia->ri_pd && !IS_ERR(ia->ri_pd))
504 		ib_dealloc_pd(ia->ri_pd);
505 	ia->ri_pd = NULL;
506 }
507 
508 /*
509  * Create unconnected endpoint.
510  */
511 int
512 rpcrdma_ep_create(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia,
513 		  struct rpcrdma_create_data_internal *cdata)
514 {
515 	struct rpcrdma_connect_private *pmsg = &ep->rep_cm_private;
516 	unsigned int max_qp_wr, max_sge;
517 	struct ib_cq *sendcq, *recvcq;
518 	int rc;
519 
520 	max_sge = min_t(unsigned int, ia->ri_device->attrs.max_sge,
521 			RPCRDMA_MAX_SEND_SGES);
522 	if (max_sge < RPCRDMA_MIN_SEND_SGES) {
523 		pr_warn("rpcrdma: HCA provides only %d send SGEs\n", max_sge);
524 		return -ENOMEM;
525 	}
526 	ia->ri_max_send_sges = max_sge - RPCRDMA_MIN_SEND_SGES;
527 
528 	if (ia->ri_device->attrs.max_qp_wr <= RPCRDMA_BACKWARD_WRS) {
529 		dprintk("RPC:       %s: insufficient wqe's available\n",
530 			__func__);
531 		return -ENOMEM;
532 	}
533 	max_qp_wr = ia->ri_device->attrs.max_qp_wr - RPCRDMA_BACKWARD_WRS - 1;
534 
535 	/* check provider's send/recv wr limits */
536 	if (cdata->max_requests > max_qp_wr)
537 		cdata->max_requests = max_qp_wr;
538 
539 	ep->rep_attr.event_handler = rpcrdma_qp_async_error_upcall;
540 	ep->rep_attr.qp_context = ep;
541 	ep->rep_attr.srq = NULL;
542 	ep->rep_attr.cap.max_send_wr = cdata->max_requests;
543 	ep->rep_attr.cap.max_send_wr += RPCRDMA_BACKWARD_WRS;
544 	ep->rep_attr.cap.max_send_wr += 1;	/* drain cqe */
545 	rc = ia->ri_ops->ro_open(ia, ep, cdata);
546 	if (rc)
547 		return rc;
548 	ep->rep_attr.cap.max_recv_wr = cdata->max_requests;
549 	ep->rep_attr.cap.max_recv_wr += RPCRDMA_BACKWARD_WRS;
550 	ep->rep_attr.cap.max_recv_wr += 1;	/* drain cqe */
551 	ep->rep_attr.cap.max_send_sge = max_sge;
552 	ep->rep_attr.cap.max_recv_sge = 1;
553 	ep->rep_attr.cap.max_inline_data = 0;
554 	ep->rep_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
555 	ep->rep_attr.qp_type = IB_QPT_RC;
556 	ep->rep_attr.port_num = ~0;
557 
558 	dprintk("RPC:       %s: requested max: dtos: send %d recv %d; "
559 		"iovs: send %d recv %d\n",
560 		__func__,
561 		ep->rep_attr.cap.max_send_wr,
562 		ep->rep_attr.cap.max_recv_wr,
563 		ep->rep_attr.cap.max_send_sge,
564 		ep->rep_attr.cap.max_recv_sge);
565 
566 	/* set trigger for requesting send completion */
567 	ep->rep_cqinit = ep->rep_attr.cap.max_send_wr/2 - 1;
568 	if (ep->rep_cqinit <= 2)
569 		ep->rep_cqinit = 0;	/* always signal? */
570 	rpcrdma_init_cqcount(ep, 0);
571 	init_waitqueue_head(&ep->rep_connect_wait);
572 	INIT_DELAYED_WORK(&ep->rep_connect_worker, rpcrdma_connect_worker);
573 
574 	sendcq = ib_alloc_cq(ia->ri_device, NULL,
575 			     ep->rep_attr.cap.max_send_wr + 1,
576 			     0, IB_POLL_SOFTIRQ);
577 	if (IS_ERR(sendcq)) {
578 		rc = PTR_ERR(sendcq);
579 		dprintk("RPC:       %s: failed to create send CQ: %i\n",
580 			__func__, rc);
581 		goto out1;
582 	}
583 
584 	recvcq = ib_alloc_cq(ia->ri_device, NULL,
585 			     ep->rep_attr.cap.max_recv_wr + 1,
586 			     0, IB_POLL_SOFTIRQ);
587 	if (IS_ERR(recvcq)) {
588 		rc = PTR_ERR(recvcq);
589 		dprintk("RPC:       %s: failed to create recv CQ: %i\n",
590 			__func__, rc);
591 		goto out2;
592 	}
593 
594 	ep->rep_attr.send_cq = sendcq;
595 	ep->rep_attr.recv_cq = recvcq;
596 
597 	/* Initialize cma parameters */
598 	memset(&ep->rep_remote_cma, 0, sizeof(ep->rep_remote_cma));
599 
600 	/* Prepare RDMA-CM private message */
601 	pmsg->cp_magic = rpcrdma_cmp_magic;
602 	pmsg->cp_version = RPCRDMA_CMP_VERSION;
603 	pmsg->cp_flags |= ia->ri_ops->ro_send_w_inv_ok;
604 	pmsg->cp_send_size = rpcrdma_encode_buffer_size(cdata->inline_wsize);
605 	pmsg->cp_recv_size = rpcrdma_encode_buffer_size(cdata->inline_rsize);
606 	ep->rep_remote_cma.private_data = pmsg;
607 	ep->rep_remote_cma.private_data_len = sizeof(*pmsg);
608 
609 	/* Client offers RDMA Read but does not initiate */
610 	ep->rep_remote_cma.initiator_depth = 0;
611 	if (ia->ri_device->attrs.max_qp_rd_atom > 32)	/* arbitrary but <= 255 */
612 		ep->rep_remote_cma.responder_resources = 32;
613 	else
614 		ep->rep_remote_cma.responder_resources =
615 						ia->ri_device->attrs.max_qp_rd_atom;
616 
617 	/* Limit transport retries so client can detect server
618 	 * GID changes quickly. RPC layer handles re-establishing
619 	 * transport connection and retransmission.
620 	 */
621 	ep->rep_remote_cma.retry_count = 6;
622 
623 	/* RPC-over-RDMA handles its own flow control. In addition,
624 	 * make all RNR NAKs visible so we know that RPC-over-RDMA
625 	 * flow control is working correctly (no NAKs should be seen).
626 	 */
627 	ep->rep_remote_cma.flow_control = 0;
628 	ep->rep_remote_cma.rnr_retry_count = 0;
629 
630 	return 0;
631 
632 out2:
633 	ib_free_cq(sendcq);
634 out1:
635 	return rc;
636 }
637 
638 /*
639  * rpcrdma_ep_destroy
640  *
641  * Disconnect and destroy endpoint. After this, the only
642  * valid operations on the ep are to free it (if dynamically
643  * allocated) or re-create it.
644  */
645 void
646 rpcrdma_ep_destroy(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia)
647 {
648 	dprintk("RPC:       %s: entering, connected is %d\n",
649 		__func__, ep->rep_connected);
650 
651 	cancel_delayed_work_sync(&ep->rep_connect_worker);
652 
653 	if (ia->ri_id->qp) {
654 		rpcrdma_ep_disconnect(ep, ia);
655 		rdma_destroy_qp(ia->ri_id);
656 		ia->ri_id->qp = NULL;
657 	}
658 
659 	ib_free_cq(ep->rep_attr.recv_cq);
660 	ib_free_cq(ep->rep_attr.send_cq);
661 }
662 
663 /* Re-establish a connection after a device removal event.
664  * Unlike a normal reconnection, a fresh PD and a new set
665  * of MRs and buffers is needed.
666  */
667 static int
668 rpcrdma_ep_recreate_xprt(struct rpcrdma_xprt *r_xprt,
669 			 struct rpcrdma_ep *ep, struct rpcrdma_ia *ia)
670 {
671 	struct sockaddr *sap = (struct sockaddr *)&r_xprt->rx_data.addr;
672 	int rc, err;
673 
674 	pr_info("%s: r_xprt = %p\n", __func__, r_xprt);
675 
676 	rc = -EHOSTUNREACH;
677 	if (rpcrdma_ia_open(r_xprt, sap))
678 		goto out1;
679 
680 	rc = -ENOMEM;
681 	err = rpcrdma_ep_create(ep, ia, &r_xprt->rx_data);
682 	if (err) {
683 		pr_err("rpcrdma: rpcrdma_ep_create returned %d\n", err);
684 		goto out2;
685 	}
686 
687 	rc = -ENETUNREACH;
688 	err = rdma_create_qp(ia->ri_id, ia->ri_pd, &ep->rep_attr);
689 	if (err) {
690 		pr_err("rpcrdma: rdma_create_qp returned %d\n", err);
691 		goto out3;
692 	}
693 
694 	rpcrdma_create_mrs(r_xprt);
695 	return 0;
696 
697 out3:
698 	rpcrdma_ep_destroy(ep, ia);
699 out2:
700 	rpcrdma_ia_close(ia);
701 out1:
702 	return rc;
703 }
704 
705 static int
706 rpcrdma_ep_reconnect(struct rpcrdma_xprt *r_xprt, struct rpcrdma_ep *ep,
707 		     struct rpcrdma_ia *ia)
708 {
709 	struct sockaddr *sap = (struct sockaddr *)&r_xprt->rx_data.addr;
710 	struct rdma_cm_id *id, *old;
711 	int err, rc;
712 
713 	dprintk("RPC:       %s: reconnecting...\n", __func__);
714 
715 	rpcrdma_ep_disconnect(ep, ia);
716 
717 	rc = -EHOSTUNREACH;
718 	id = rpcrdma_create_id(r_xprt, ia, sap);
719 	if (IS_ERR(id))
720 		goto out;
721 
722 	/* As long as the new ID points to the same device as the
723 	 * old ID, we can reuse the transport's existing PD and all
724 	 * previously allocated MRs. Also, the same device means
725 	 * the transport's previous DMA mappings are still valid.
726 	 *
727 	 * This is a sanity check only. There should be no way these
728 	 * point to two different devices here.
729 	 */
730 	old = id;
731 	rc = -ENETUNREACH;
732 	if (ia->ri_device != id->device) {
733 		pr_err("rpcrdma: can't reconnect on different device!\n");
734 		goto out_destroy;
735 	}
736 
737 	err = rdma_create_qp(id, ia->ri_pd, &ep->rep_attr);
738 	if (err) {
739 		dprintk("RPC:       %s: rdma_create_qp returned %d\n",
740 			__func__, err);
741 		goto out_destroy;
742 	}
743 
744 	/* Atomically replace the transport's ID and QP. */
745 	rc = 0;
746 	old = ia->ri_id;
747 	ia->ri_id = id;
748 	rdma_destroy_qp(old);
749 
750 out_destroy:
751 	rdma_destroy_id(old);
752 out:
753 	return rc;
754 }
755 
756 /*
757  * Connect unconnected endpoint.
758  */
759 int
760 rpcrdma_ep_connect(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia)
761 {
762 	struct rpcrdma_xprt *r_xprt = container_of(ia, struct rpcrdma_xprt,
763 						   rx_ia);
764 	unsigned int extras;
765 	int rc;
766 
767 retry:
768 	switch (ep->rep_connected) {
769 	case 0:
770 		dprintk("RPC:       %s: connecting...\n", __func__);
771 		rc = rdma_create_qp(ia->ri_id, ia->ri_pd, &ep->rep_attr);
772 		if (rc) {
773 			dprintk("RPC:       %s: rdma_create_qp failed %i\n",
774 				__func__, rc);
775 			rc = -ENETUNREACH;
776 			goto out_noupdate;
777 		}
778 		break;
779 	case -ENODEV:
780 		rc = rpcrdma_ep_recreate_xprt(r_xprt, ep, ia);
781 		if (rc)
782 			goto out_noupdate;
783 		break;
784 	default:
785 		rc = rpcrdma_ep_reconnect(r_xprt, ep, ia);
786 		if (rc)
787 			goto out;
788 	}
789 
790 	ep->rep_connected = 0;
791 
792 	rc = rdma_connect(ia->ri_id, &ep->rep_remote_cma);
793 	if (rc) {
794 		dprintk("RPC:       %s: rdma_connect() failed with %i\n",
795 				__func__, rc);
796 		goto out;
797 	}
798 
799 	wait_event_interruptible(ep->rep_connect_wait, ep->rep_connected != 0);
800 	if (ep->rep_connected <= 0) {
801 		if (ep->rep_connected == -EAGAIN)
802 			goto retry;
803 		rc = ep->rep_connected;
804 		goto out;
805 	}
806 
807 	dprintk("RPC:       %s: connected\n", __func__);
808 	extras = r_xprt->rx_buf.rb_bc_srv_max_requests;
809 	if (extras)
810 		rpcrdma_ep_post_extra_recv(r_xprt, extras);
811 
812 out:
813 	if (rc)
814 		ep->rep_connected = rc;
815 
816 out_noupdate:
817 	return rc;
818 }
819 
820 /*
821  * rpcrdma_ep_disconnect
822  *
823  * This is separate from destroy to facilitate the ability
824  * to reconnect without recreating the endpoint.
825  *
826  * This call is not reentrant, and must not be made in parallel
827  * on the same endpoint.
828  */
829 void
830 rpcrdma_ep_disconnect(struct rpcrdma_ep *ep, struct rpcrdma_ia *ia)
831 {
832 	int rc;
833 
834 	rc = rdma_disconnect(ia->ri_id);
835 	if (!rc) {
836 		/* returns without wait if not connected */
837 		wait_event_interruptible(ep->rep_connect_wait,
838 							ep->rep_connected != 1);
839 		dprintk("RPC:       %s: after wait, %sconnected\n", __func__,
840 			(ep->rep_connected == 1) ? "still " : "dis");
841 	} else {
842 		dprintk("RPC:       %s: rdma_disconnect %i\n", __func__, rc);
843 		ep->rep_connected = rc;
844 	}
845 
846 	ib_drain_qp(ia->ri_id->qp);
847 }
848 
849 static void
850 rpcrdma_mr_recovery_worker(struct work_struct *work)
851 {
852 	struct rpcrdma_buffer *buf = container_of(work, struct rpcrdma_buffer,
853 						  rb_recovery_worker.work);
854 	struct rpcrdma_mw *mw;
855 
856 	spin_lock(&buf->rb_recovery_lock);
857 	while (!list_empty(&buf->rb_stale_mrs)) {
858 		mw = rpcrdma_pop_mw(&buf->rb_stale_mrs);
859 		spin_unlock(&buf->rb_recovery_lock);
860 
861 		dprintk("RPC:       %s: recovering MR %p\n", __func__, mw);
862 		mw->mw_xprt->rx_ia.ri_ops->ro_recover_mr(mw);
863 
864 		spin_lock(&buf->rb_recovery_lock);
865 	}
866 	spin_unlock(&buf->rb_recovery_lock);
867 }
868 
869 void
870 rpcrdma_defer_mr_recovery(struct rpcrdma_mw *mw)
871 {
872 	struct rpcrdma_xprt *r_xprt = mw->mw_xprt;
873 	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
874 
875 	spin_lock(&buf->rb_recovery_lock);
876 	rpcrdma_push_mw(mw, &buf->rb_stale_mrs);
877 	spin_unlock(&buf->rb_recovery_lock);
878 
879 	schedule_delayed_work(&buf->rb_recovery_worker, 0);
880 }
881 
882 static void
883 rpcrdma_create_mrs(struct rpcrdma_xprt *r_xprt)
884 {
885 	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
886 	struct rpcrdma_ia *ia = &r_xprt->rx_ia;
887 	unsigned int count;
888 	LIST_HEAD(free);
889 	LIST_HEAD(all);
890 
891 	for (count = 0; count < 32; count++) {
892 		struct rpcrdma_mw *mw;
893 		int rc;
894 
895 		mw = kzalloc(sizeof(*mw), GFP_KERNEL);
896 		if (!mw)
897 			break;
898 
899 		rc = ia->ri_ops->ro_init_mr(ia, mw);
900 		if (rc) {
901 			kfree(mw);
902 			break;
903 		}
904 
905 		mw->mw_xprt = r_xprt;
906 
907 		list_add(&mw->mw_list, &free);
908 		list_add(&mw->mw_all, &all);
909 	}
910 
911 	spin_lock(&buf->rb_mwlock);
912 	list_splice(&free, &buf->rb_mws);
913 	list_splice(&all, &buf->rb_all);
914 	r_xprt->rx_stats.mrs_allocated += count;
915 	spin_unlock(&buf->rb_mwlock);
916 
917 	dprintk("RPC:       %s: created %u MRs\n", __func__, count);
918 }
919 
920 static void
921 rpcrdma_mr_refresh_worker(struct work_struct *work)
922 {
923 	struct rpcrdma_buffer *buf = container_of(work, struct rpcrdma_buffer,
924 						  rb_refresh_worker.work);
925 	struct rpcrdma_xprt *r_xprt = container_of(buf, struct rpcrdma_xprt,
926 						   rx_buf);
927 
928 	rpcrdma_create_mrs(r_xprt);
929 }
930 
931 struct rpcrdma_req *
932 rpcrdma_create_req(struct rpcrdma_xprt *r_xprt)
933 {
934 	struct rpcrdma_buffer *buffer = &r_xprt->rx_buf;
935 	struct rpcrdma_req *req;
936 
937 	req = kzalloc(sizeof(*req), GFP_KERNEL);
938 	if (req == NULL)
939 		return ERR_PTR(-ENOMEM);
940 
941 	spin_lock(&buffer->rb_reqslock);
942 	list_add(&req->rl_all, &buffer->rb_allreqs);
943 	spin_unlock(&buffer->rb_reqslock);
944 	req->rl_cqe.done = rpcrdma_wc_send;
945 	req->rl_buffer = &r_xprt->rx_buf;
946 	INIT_LIST_HEAD(&req->rl_registered);
947 	req->rl_send_wr.next = NULL;
948 	req->rl_send_wr.wr_cqe = &req->rl_cqe;
949 	req->rl_send_wr.sg_list = req->rl_send_sge;
950 	req->rl_send_wr.opcode = IB_WR_SEND;
951 	return req;
952 }
953 
954 struct rpcrdma_rep *
955 rpcrdma_create_rep(struct rpcrdma_xprt *r_xprt)
956 {
957 	struct rpcrdma_create_data_internal *cdata = &r_xprt->rx_data;
958 	struct rpcrdma_rep *rep;
959 	int rc;
960 
961 	rc = -ENOMEM;
962 	rep = kzalloc(sizeof(*rep), GFP_KERNEL);
963 	if (rep == NULL)
964 		goto out;
965 
966 	rep->rr_rdmabuf = rpcrdma_alloc_regbuf(cdata->inline_rsize,
967 					       DMA_FROM_DEVICE, GFP_KERNEL);
968 	if (IS_ERR(rep->rr_rdmabuf)) {
969 		rc = PTR_ERR(rep->rr_rdmabuf);
970 		goto out_free;
971 	}
972 	xdr_buf_init(&rep->rr_hdrbuf, rep->rr_rdmabuf->rg_base,
973 		     rdmab_length(rep->rr_rdmabuf));
974 
975 	rep->rr_cqe.done = rpcrdma_wc_receive;
976 	rep->rr_rxprt = r_xprt;
977 	INIT_WORK(&rep->rr_work, rpcrdma_reply_handler);
978 	rep->rr_recv_wr.next = NULL;
979 	rep->rr_recv_wr.wr_cqe = &rep->rr_cqe;
980 	rep->rr_recv_wr.sg_list = &rep->rr_rdmabuf->rg_iov;
981 	rep->rr_recv_wr.num_sge = 1;
982 	return rep;
983 
984 out_free:
985 	kfree(rep);
986 out:
987 	return ERR_PTR(rc);
988 }
989 
990 int
991 rpcrdma_buffer_create(struct rpcrdma_xprt *r_xprt)
992 {
993 	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
994 	int i, rc;
995 
996 	buf->rb_max_requests = r_xprt->rx_data.max_requests;
997 	buf->rb_bc_srv_max_requests = 0;
998 	atomic_set(&buf->rb_credits, 1);
999 	spin_lock_init(&buf->rb_mwlock);
1000 	spin_lock_init(&buf->rb_lock);
1001 	spin_lock_init(&buf->rb_recovery_lock);
1002 	INIT_LIST_HEAD(&buf->rb_mws);
1003 	INIT_LIST_HEAD(&buf->rb_all);
1004 	INIT_LIST_HEAD(&buf->rb_stale_mrs);
1005 	INIT_DELAYED_WORK(&buf->rb_refresh_worker,
1006 			  rpcrdma_mr_refresh_worker);
1007 	INIT_DELAYED_WORK(&buf->rb_recovery_worker,
1008 			  rpcrdma_mr_recovery_worker);
1009 
1010 	rpcrdma_create_mrs(r_xprt);
1011 
1012 	INIT_LIST_HEAD(&buf->rb_send_bufs);
1013 	INIT_LIST_HEAD(&buf->rb_allreqs);
1014 	spin_lock_init(&buf->rb_reqslock);
1015 	for (i = 0; i < buf->rb_max_requests; i++) {
1016 		struct rpcrdma_req *req;
1017 
1018 		req = rpcrdma_create_req(r_xprt);
1019 		if (IS_ERR(req)) {
1020 			dprintk("RPC:       %s: request buffer %d alloc"
1021 				" failed\n", __func__, i);
1022 			rc = PTR_ERR(req);
1023 			goto out;
1024 		}
1025 		req->rl_backchannel = false;
1026 		list_add(&req->rl_list, &buf->rb_send_bufs);
1027 	}
1028 
1029 	INIT_LIST_HEAD(&buf->rb_recv_bufs);
1030 	for (i = 0; i < buf->rb_max_requests + RPCRDMA_MAX_BC_REQUESTS; i++) {
1031 		struct rpcrdma_rep *rep;
1032 
1033 		rep = rpcrdma_create_rep(r_xprt);
1034 		if (IS_ERR(rep)) {
1035 			dprintk("RPC:       %s: reply buffer %d alloc failed\n",
1036 				__func__, i);
1037 			rc = PTR_ERR(rep);
1038 			goto out;
1039 		}
1040 		list_add(&rep->rr_list, &buf->rb_recv_bufs);
1041 	}
1042 
1043 	return 0;
1044 out:
1045 	rpcrdma_buffer_destroy(buf);
1046 	return rc;
1047 }
1048 
1049 static struct rpcrdma_req *
1050 rpcrdma_buffer_get_req_locked(struct rpcrdma_buffer *buf)
1051 {
1052 	struct rpcrdma_req *req;
1053 
1054 	req = list_first_entry(&buf->rb_send_bufs,
1055 			       struct rpcrdma_req, rl_list);
1056 	list_del_init(&req->rl_list);
1057 	return req;
1058 }
1059 
1060 static struct rpcrdma_rep *
1061 rpcrdma_buffer_get_rep_locked(struct rpcrdma_buffer *buf)
1062 {
1063 	struct rpcrdma_rep *rep;
1064 
1065 	rep = list_first_entry(&buf->rb_recv_bufs,
1066 			       struct rpcrdma_rep, rr_list);
1067 	list_del(&rep->rr_list);
1068 	return rep;
1069 }
1070 
1071 static void
1072 rpcrdma_destroy_rep(struct rpcrdma_rep *rep)
1073 {
1074 	rpcrdma_free_regbuf(rep->rr_rdmabuf);
1075 	kfree(rep);
1076 }
1077 
1078 void
1079 rpcrdma_destroy_req(struct rpcrdma_req *req)
1080 {
1081 	rpcrdma_free_regbuf(req->rl_recvbuf);
1082 	rpcrdma_free_regbuf(req->rl_sendbuf);
1083 	rpcrdma_free_regbuf(req->rl_rdmabuf);
1084 	kfree(req);
1085 }
1086 
1087 static void
1088 rpcrdma_destroy_mrs(struct rpcrdma_buffer *buf)
1089 {
1090 	struct rpcrdma_xprt *r_xprt = container_of(buf, struct rpcrdma_xprt,
1091 						   rx_buf);
1092 	struct rpcrdma_ia *ia = rdmab_to_ia(buf);
1093 	struct rpcrdma_mw *mw;
1094 	unsigned int count;
1095 
1096 	count = 0;
1097 	spin_lock(&buf->rb_mwlock);
1098 	while (!list_empty(&buf->rb_all)) {
1099 		mw = list_entry(buf->rb_all.next, struct rpcrdma_mw, mw_all);
1100 		list_del(&mw->mw_all);
1101 
1102 		spin_unlock(&buf->rb_mwlock);
1103 		ia->ri_ops->ro_release_mr(mw);
1104 		count++;
1105 		spin_lock(&buf->rb_mwlock);
1106 	}
1107 	spin_unlock(&buf->rb_mwlock);
1108 	r_xprt->rx_stats.mrs_allocated = 0;
1109 
1110 	dprintk("RPC:       %s: released %u MRs\n", __func__, count);
1111 }
1112 
1113 void
1114 rpcrdma_buffer_destroy(struct rpcrdma_buffer *buf)
1115 {
1116 	cancel_delayed_work_sync(&buf->rb_recovery_worker);
1117 	cancel_delayed_work_sync(&buf->rb_refresh_worker);
1118 
1119 	while (!list_empty(&buf->rb_recv_bufs)) {
1120 		struct rpcrdma_rep *rep;
1121 
1122 		rep = rpcrdma_buffer_get_rep_locked(buf);
1123 		rpcrdma_destroy_rep(rep);
1124 	}
1125 	buf->rb_send_count = 0;
1126 
1127 	spin_lock(&buf->rb_reqslock);
1128 	while (!list_empty(&buf->rb_allreqs)) {
1129 		struct rpcrdma_req *req;
1130 
1131 		req = list_first_entry(&buf->rb_allreqs,
1132 				       struct rpcrdma_req, rl_all);
1133 		list_del(&req->rl_all);
1134 
1135 		spin_unlock(&buf->rb_reqslock);
1136 		rpcrdma_destroy_req(req);
1137 		spin_lock(&buf->rb_reqslock);
1138 	}
1139 	spin_unlock(&buf->rb_reqslock);
1140 	buf->rb_recv_count = 0;
1141 
1142 	rpcrdma_destroy_mrs(buf);
1143 }
1144 
1145 struct rpcrdma_mw *
1146 rpcrdma_get_mw(struct rpcrdma_xprt *r_xprt)
1147 {
1148 	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1149 	struct rpcrdma_mw *mw = NULL;
1150 
1151 	spin_lock(&buf->rb_mwlock);
1152 	if (!list_empty(&buf->rb_mws))
1153 		mw = rpcrdma_pop_mw(&buf->rb_mws);
1154 	spin_unlock(&buf->rb_mwlock);
1155 
1156 	if (!mw)
1157 		goto out_nomws;
1158 	mw->mw_flags = 0;
1159 	return mw;
1160 
1161 out_nomws:
1162 	dprintk("RPC:       %s: no MWs available\n", __func__);
1163 	if (r_xprt->rx_ep.rep_connected != -ENODEV)
1164 		schedule_delayed_work(&buf->rb_refresh_worker, 0);
1165 
1166 	/* Allow the reply handler and refresh worker to run */
1167 	cond_resched();
1168 
1169 	return NULL;
1170 }
1171 
1172 void
1173 rpcrdma_put_mw(struct rpcrdma_xprt *r_xprt, struct rpcrdma_mw *mw)
1174 {
1175 	struct rpcrdma_buffer *buf = &r_xprt->rx_buf;
1176 
1177 	spin_lock(&buf->rb_mwlock);
1178 	rpcrdma_push_mw(mw, &buf->rb_mws);
1179 	spin_unlock(&buf->rb_mwlock);
1180 }
1181 
1182 static struct rpcrdma_rep *
1183 rpcrdma_buffer_get_rep(struct rpcrdma_buffer *buffers)
1184 {
1185 	/* If an RPC previously completed without a reply (say, a
1186 	 * credential problem or a soft timeout occurs) then hold off
1187 	 * on supplying more Receive buffers until the number of new
1188 	 * pending RPCs catches up to the number of posted Receives.
1189 	 */
1190 	if (unlikely(buffers->rb_send_count < buffers->rb_recv_count))
1191 		return NULL;
1192 
1193 	if (unlikely(list_empty(&buffers->rb_recv_bufs)))
1194 		return NULL;
1195 	buffers->rb_recv_count++;
1196 	return rpcrdma_buffer_get_rep_locked(buffers);
1197 }
1198 
1199 /*
1200  * Get a set of request/reply buffers.
1201  *
1202  * Reply buffer (if available) is attached to send buffer upon return.
1203  */
1204 struct rpcrdma_req *
1205 rpcrdma_buffer_get(struct rpcrdma_buffer *buffers)
1206 {
1207 	struct rpcrdma_req *req;
1208 
1209 	spin_lock(&buffers->rb_lock);
1210 	if (list_empty(&buffers->rb_send_bufs))
1211 		goto out_reqbuf;
1212 	buffers->rb_send_count++;
1213 	req = rpcrdma_buffer_get_req_locked(buffers);
1214 	req->rl_reply = rpcrdma_buffer_get_rep(buffers);
1215 	spin_unlock(&buffers->rb_lock);
1216 	return req;
1217 
1218 out_reqbuf:
1219 	spin_unlock(&buffers->rb_lock);
1220 	pr_warn("RPC:       %s: out of request buffers\n", __func__);
1221 	return NULL;
1222 }
1223 
1224 /*
1225  * Put request/reply buffers back into pool.
1226  * Pre-decrement counter/array index.
1227  */
1228 void
1229 rpcrdma_buffer_put(struct rpcrdma_req *req)
1230 {
1231 	struct rpcrdma_buffer *buffers = req->rl_buffer;
1232 	struct rpcrdma_rep *rep = req->rl_reply;
1233 
1234 	req->rl_send_wr.num_sge = 0;
1235 	req->rl_reply = NULL;
1236 
1237 	spin_lock(&buffers->rb_lock);
1238 	buffers->rb_send_count--;
1239 	list_add_tail(&req->rl_list, &buffers->rb_send_bufs);
1240 	if (rep) {
1241 		buffers->rb_recv_count--;
1242 		list_add_tail(&rep->rr_list, &buffers->rb_recv_bufs);
1243 	}
1244 	spin_unlock(&buffers->rb_lock);
1245 }
1246 
1247 /*
1248  * Recover reply buffers from pool.
1249  * This happens when recovering from disconnect.
1250  */
1251 void
1252 rpcrdma_recv_buffer_get(struct rpcrdma_req *req)
1253 {
1254 	struct rpcrdma_buffer *buffers = req->rl_buffer;
1255 
1256 	spin_lock(&buffers->rb_lock);
1257 	req->rl_reply = rpcrdma_buffer_get_rep(buffers);
1258 	spin_unlock(&buffers->rb_lock);
1259 }
1260 
1261 /*
1262  * Put reply buffers back into pool when not attached to
1263  * request. This happens in error conditions.
1264  */
1265 void
1266 rpcrdma_recv_buffer_put(struct rpcrdma_rep *rep)
1267 {
1268 	struct rpcrdma_buffer *buffers = &rep->rr_rxprt->rx_buf;
1269 
1270 	spin_lock(&buffers->rb_lock);
1271 	buffers->rb_recv_count--;
1272 	list_add_tail(&rep->rr_list, &buffers->rb_recv_bufs);
1273 	spin_unlock(&buffers->rb_lock);
1274 }
1275 
1276 /**
1277  * rpcrdma_alloc_regbuf - allocate and DMA-map memory for SEND/RECV buffers
1278  * @size: size of buffer to be allocated, in bytes
1279  * @direction: direction of data movement
1280  * @flags: GFP flags
1281  *
1282  * Returns an ERR_PTR, or a pointer to a regbuf, a buffer that
1283  * can be persistently DMA-mapped for I/O.
1284  *
1285  * xprtrdma uses a regbuf for posting an outgoing RDMA SEND, or for
1286  * receiving the payload of RDMA RECV operations. During Long Calls
1287  * or Replies they may be registered externally via ro_map.
1288  */
1289 struct rpcrdma_regbuf *
1290 rpcrdma_alloc_regbuf(size_t size, enum dma_data_direction direction,
1291 		     gfp_t flags)
1292 {
1293 	struct rpcrdma_regbuf *rb;
1294 
1295 	rb = kmalloc(sizeof(*rb) + size, flags);
1296 	if (rb == NULL)
1297 		return ERR_PTR(-ENOMEM);
1298 
1299 	rb->rg_device = NULL;
1300 	rb->rg_direction = direction;
1301 	rb->rg_iov.length = size;
1302 
1303 	return rb;
1304 }
1305 
1306 /**
1307  * __rpcrdma_map_regbuf - DMA-map a regbuf
1308  * @ia: controlling rpcrdma_ia
1309  * @rb: regbuf to be mapped
1310  */
1311 bool
1312 __rpcrdma_dma_map_regbuf(struct rpcrdma_ia *ia, struct rpcrdma_regbuf *rb)
1313 {
1314 	struct ib_device *device = ia->ri_device;
1315 
1316 	if (rb->rg_direction == DMA_NONE)
1317 		return false;
1318 
1319 	rb->rg_iov.addr = ib_dma_map_single(device,
1320 					    (void *)rb->rg_base,
1321 					    rdmab_length(rb),
1322 					    rb->rg_direction);
1323 	if (ib_dma_mapping_error(device, rdmab_addr(rb)))
1324 		return false;
1325 
1326 	rb->rg_device = device;
1327 	rb->rg_iov.lkey = ia->ri_pd->local_dma_lkey;
1328 	return true;
1329 }
1330 
1331 static void
1332 rpcrdma_dma_unmap_regbuf(struct rpcrdma_regbuf *rb)
1333 {
1334 	if (!rpcrdma_regbuf_is_mapped(rb))
1335 		return;
1336 
1337 	ib_dma_unmap_single(rb->rg_device, rdmab_addr(rb),
1338 			    rdmab_length(rb), rb->rg_direction);
1339 	rb->rg_device = NULL;
1340 }
1341 
1342 /**
1343  * rpcrdma_free_regbuf - deregister and free registered buffer
1344  * @rb: regbuf to be deregistered and freed
1345  */
1346 void
1347 rpcrdma_free_regbuf(struct rpcrdma_regbuf *rb)
1348 {
1349 	if (!rb)
1350 		return;
1351 
1352 	rpcrdma_dma_unmap_regbuf(rb);
1353 	kfree(rb);
1354 }
1355 
1356 /*
1357  * Prepost any receive buffer, then post send.
1358  *
1359  * Receive buffer is donated to hardware, reclaimed upon recv completion.
1360  */
1361 int
1362 rpcrdma_ep_post(struct rpcrdma_ia *ia,
1363 		struct rpcrdma_ep *ep,
1364 		struct rpcrdma_req *req)
1365 {
1366 	struct ib_send_wr *send_wr = &req->rl_send_wr;
1367 	struct ib_send_wr *send_wr_fail;
1368 	int rc;
1369 
1370 	if (req->rl_reply) {
1371 		rc = rpcrdma_ep_post_recv(ia, req->rl_reply);
1372 		if (rc)
1373 			return rc;
1374 		req->rl_reply = NULL;
1375 	}
1376 
1377 	dprintk("RPC:       %s: posting %d s/g entries\n",
1378 		__func__, send_wr->num_sge);
1379 
1380 	rpcrdma_set_signaled(ep, send_wr);
1381 	rc = ib_post_send(ia->ri_id->qp, send_wr, &send_wr_fail);
1382 	if (rc)
1383 		goto out_postsend_err;
1384 	return 0;
1385 
1386 out_postsend_err:
1387 	pr_err("rpcrdma: RDMA Send ib_post_send returned %i\n", rc);
1388 	return -ENOTCONN;
1389 }
1390 
1391 int
1392 rpcrdma_ep_post_recv(struct rpcrdma_ia *ia,
1393 		     struct rpcrdma_rep *rep)
1394 {
1395 	struct ib_recv_wr *recv_wr_fail;
1396 	int rc;
1397 
1398 	if (!rpcrdma_dma_map_regbuf(ia, rep->rr_rdmabuf))
1399 		goto out_map;
1400 	rc = ib_post_recv(ia->ri_id->qp, &rep->rr_recv_wr, &recv_wr_fail);
1401 	if (rc)
1402 		goto out_postrecv;
1403 	return 0;
1404 
1405 out_map:
1406 	pr_err("rpcrdma: failed to DMA map the Receive buffer\n");
1407 	return -EIO;
1408 
1409 out_postrecv:
1410 	pr_err("rpcrdma: ib_post_recv returned %i\n", rc);
1411 	return -ENOTCONN;
1412 }
1413 
1414 /**
1415  * rpcrdma_ep_post_extra_recv - Post buffers for incoming backchannel requests
1416  * @r_xprt: transport associated with these backchannel resources
1417  * @min_reqs: minimum number of incoming requests expected
1418  *
1419  * Returns zero if all requested buffers were posted, or a negative errno.
1420  */
1421 int
1422 rpcrdma_ep_post_extra_recv(struct rpcrdma_xprt *r_xprt, unsigned int count)
1423 {
1424 	struct rpcrdma_buffer *buffers = &r_xprt->rx_buf;
1425 	struct rpcrdma_ia *ia = &r_xprt->rx_ia;
1426 	struct rpcrdma_rep *rep;
1427 	int rc;
1428 
1429 	while (count--) {
1430 		spin_lock(&buffers->rb_lock);
1431 		if (list_empty(&buffers->rb_recv_bufs))
1432 			goto out_reqbuf;
1433 		rep = rpcrdma_buffer_get_rep_locked(buffers);
1434 		spin_unlock(&buffers->rb_lock);
1435 
1436 		rc = rpcrdma_ep_post_recv(ia, rep);
1437 		if (rc)
1438 			goto out_rc;
1439 	}
1440 
1441 	return 0;
1442 
1443 out_reqbuf:
1444 	spin_unlock(&buffers->rb_lock);
1445 	pr_warn("%s: no extra receive buffers\n", __func__);
1446 	return -ENOMEM;
1447 
1448 out_rc:
1449 	rpcrdma_recv_buffer_put(rep);
1450 	return rc;
1451 }
1452